Goal-Reasoning in StarCraft: Brood War through Multilevel Planning

Real-Time Strategy video-games (RTSVGs) are challenging for most deliberative approaches, such as Automated Planning. This is due to (i) the dynamic changes of the environment; (ii) and the wide variety of potential actions that can be performed over the environment. The aim is “to win the match”. Besides, RTSVGs presents an additional challenge: managing goals during the game is extremely hard. They change as the game state evolves either because of actions performed by the different agents (player and opponents), by new available information or by unexpected changes of the environment. Thus, generating a detailed sequence of actions –plan– to win the match is not effective in the long term.
In this paper, we propose an autonomous approach based on two levels of declarative Automated Planning. They are included inside a planning and execution architecture. The high-level, macromanagement, searches and suggests a set of soft-goals according to the current state and the available features of the agent. The low level, micromanagement, generates short plans of actions to reach the soft-goals generated by the high level. We claim that the ability of self-generating goals improves the plan generation and execution performance in a dynamic environment. Finally, we present a preliminary empirical evaluation of this approach tested on StarCraft: Brood War